Facial Expression Recognition Based On Bilinear Model

In recent years, research of facial expression recognition has been welldeveloped. As most of the methods are based on two-dimensional image or video, theeffect of recognition is greatly affected when the facial gesture or the lightingconditions change. With the development and popularization of three-dimensionalinformation acquisition device, it’s much more convenience to obtain3D data.3Dfacial data contains more comprehensive information about face, which has providednew approaches to solve the problem of facial expression recognition when lightingconditions and facial gesture change. However, the methods of facial expressionrecognition based on3D data need further research on aspects of the expression andextraction of the facial feature and so on.Recently, according to the difference of extraction mode, the3D facialexpression recognition methods can be generally depart as methods based on localfeatures and global features. The methods based on local features show a trend ofdiversification on the feature expression, that there is no unified representation whichhas been endorsed unanimously. The methods based on global features can expressthe global information of a human face, so that it has become a hot report topic inthese years. This paper will research on a facial expression recognition method basedon bilinear model which is based on global features.The bilinear model of the3D facial expression data contains two sets ofparameters about facial identity feature and facial expression feature. These twogroups of features represent the identity property and the expression property of ahuman face. Building bilinear model needs a unified representation of3D facialexpression data, therefore this paper complete the registration of3D point cloud datawith a registration algorithm based on Thin Plate Spline (TPS) after research andanalysis of the alignment methods. We build a symmetrical3D facial expression databilinear model based on the data after registration to get the expression feature andidentity feature of faces. For better classification of expressions, we propose aexpression classification principle based on approximate identity information. Thatmeans we choose facial expression data with similar identify features, and dosimilarity measure during those sample data, to realize the expression identification atlast. In the experiment, we choose700faces data from BU-3DFE database as thetraining set and another700faces as the test set. The result of the experiment showsthat the methods of facial expression recognition we propose in this paper can departthe identify features and the expression features well and has a higher expressionrecognition rate.